Comment by joe_the_user | Hacker News Reader

Comment by joe_the_user | Hacker News Reader

In the next five years, computer programs that can think will read legal documents and give medical advice.

Aside from the other points, taking "AI" as it exists in it's present form (deep neural networks and related) as specifically the bringer of unlimited wealth certainly puffs up the various "AI companies" notably OpenAI (It should be noted that OpenAI's most famous product, GPT-3, can generate strings that sound a lot like legal or medical advice but it so far "demonstrates non-understanding on a regular basis". Don't follow it's advice to kill yourself, for example).

It really should be said that deep learning, in particular, is still just one technology that's very good at some things, kind of impressive but not functional at other things, and just unable to do other things (actual understanding of biology, for example, seems well beyond them). I don't think this situation has changed since deep learning began it's hype cycle (which isn't to say it's "nothing", it just doesn't seem like to bring us "everything", a scenario the article literally sketches).

Automation has proceeded apace, automation in general has brought us enough resources right now to give minimal comfort to most people in the planet (as people have noted).

But automation has generally succeeded in situations where everything is controlled - ie, factories. Self-driving cars are forever five years away given the 5% or 1% or whatever level of unpredictable variable involved. Progress on robots that can interact well with either humans or "the messy real world" even in very limited terms has been painfully slow and I expect this to continue.

One other thing I'd like to add.

The scenario of AI mostly replacing people like doctors and lawyers involves bizarre paradoxes beyond whether deep learning "AI" works as advertised. Suppose you can train an "AI" to read legal papers or diagnose patients based on X-rays. That training is done from the data of real life lawyers and doctors actions. Suppose, best case scenario (very unrealistic imo btw), you have a complete "snapshot" of the behavior of lawyers and doctors in a given year. The problem is reality changes, you need new lawyering and doctoring behaviors after N years. Doctors need interpret new maladies, lawyers need to cite new decisions and both need to interpret new language forms that appear. But if you've actually removed the real lawyers and doctors, where would you get the new training data?

And this is just taking AI at it's word.

The only way you could be beguiled by this framing is if you don't understand just how inept most doctors and lawyers really are.

A future configuration will probably look something like: far fewer highly talented doctors and lawyers remain employable while the rest are replaced by AI that's shown to be vastly more capable, and that is continually enhanced by the encoded expertise of said highly talented remaining specialists.

You don't understand just how inept most doctors and lawyers really are.

It doesn't matter how competent or incompetent whatever professional might be. The only thing a deep learning application is going to do is duplicate their behavior. Deep learning involves no "thinking" at all. Just a very elaborate, brute force curve fitting. If the doctors are on average "incompetent" so will be the deep learning app (ie, you kind fall for the sort of "since it's a machine, it will be accurate" fallacy that makes people want to trust self-driving cars)

A future configuration will probably look something like: far fewer highly talented doctors and lawyers.

If you really automated the work of lawyers and doctors with explicit, maybe. BUT that isn't how "deep learning" work. Deep learning just uses data and the problem is you need sufficient data, a sufficiently large corpus of data to show by many, many examples what the thing should do.

Oddly enough, your scenario of high expert adding their expertise to the system is much more like the original Gofai model where a few experts would hypothetically program in their expertise. That scenario fell with difficulty of expertise programming. The present systems can't work that way.

> Oddly enough, your scenario of high expert adding their expertise to the system is much more like the original Gofai model where a few experts would hypothetically program in their expertise. That scenario fell with difficulty of expertise programming. The present systems can't work that way.

Taking a generalized approach will fail. Taking a tailored, domain-specific one that incrementally carves out use cases will be the basis of future success in these spaces.

> fallacy that makes people want to trust self-driving cars

A lot of successful L4+ autonomous vehicles today, contrary to what the press releases want you to believe, are architectured first and foremost as non-learning (i.e. traditional robotics) systems, with relatively well-defined domain-specific sub-problems carved out and delegated to learning-based methods (e.g. recognizing all cars/humans/signs/... in images captured by the vehicle's cameras). These problems tend to have well-defined metrics and massive real-world data sets backing them up, and are increasingly more common to report how confident they are in the provided results.

ADVs have come a long way despite all the doubt, and the top players are finally getting confident in removing the human from the driver's seat. This is not trivial in the post-Uber-ADV-fatal-accident world.

Source: I work on ADVs.

the opposite if this is true. with scifi ai on that level, bad lawyers would just pay to use the ai for legal work and focus on developing social relationships to get valuable clients. this is how a lot of professionals spend their time now

most things in life that make a lot of money are social. if everyone has access to ai knowledge and skills that can be copied, then the social aspect is even more important

I assume the author is using AI to mean software in its broader sense, not just a machine learning algorithm.

Somebody will eventually write and algorithm that actually understands the laws as they are written, the case law that interpret them, and can read and write contracts. The algorithm might not have anything to do with ML.

I would like to think one day we'll actually understand how our bodies work and can diagnose problems as if we were debugging software rather than stabbing into the dark with drugs.

Some of it is just framing. You can replace a lot of what doctors do with regular tech and plain old organization. You could have done it for decades. Problem is regulations put a lot of limitations on who can perform what service. You can get the best well-informed advice from wherever, but you still need to go to a doctor for your prescription.

I could very well see a situation where "AI" is what finally gets regulators to loosen up, but what actually gets implemented will be traditional stuff that works. It may take a "surgery robot" to allow greater freedom in training people to perform surgeries.

Won't be surprised if in 20 years self driving cars will indeed come to dominate. But most "AI" will get dropped and roads will be retrofitted with something dead simple that assists cars in navigating (I believe that idea goes back to 50s or 60s too).

In 1995, a self driving car drove 98% of the way across the country. Think what these same people predicting AI today would have predicted in 1995. They would probably believe we were 10 years away from self driving cars in every household. We still dont have a mass produced level 3 system in 2021.

https://en.m.wikipedia.org/wiki/History_of_self-driving_cars

On the other hand, the rate of change itself has a rate of change.

Thanks. I was starting to think I was the only one not believing in the AI hype train.

At least a couple of times in the article the author talks about software "that can think". Deep neural nets can't "think" at least not in the sense that humans can think. I don't think we have any software yet that's close to "thinking".

You can, today, take a self driving taxi in phoenix with no (other) human in the car.

Reply on news.ycombinator.com